Anthropic Grants METR Unprecedented Insider Access to Audit AI Safety Risks

TL;DR

• Anthropic has granted nonprofit METR employee-level access to its models, systems, and safety processes in a landmark transparency agreement.
• METR will be able to publish its findings independently, as Anthropic has waived veto power over its evaluations.
• The deal is being hailed as a potential new blueprint for third-party oversight of frontier AI labs.

A New Standard for Frontier Oversight

Anthropic is opening its doors to outside scrutiny in a way no frontier AI lab has before. The company has struck a landmark agreement giving the nonprofit AI evaluation organization METR unprecedented, employee-level access to evaluate its models, risks, and safety practices.

The arrangement goes far beyond the typical pre-deployment API access granted to external testers. METR researchers will be embedded much like internal staff, with deep visibility into Anthropic's development pipeline, safety testing, and deployment decisions.

Inside the Unprecedented Access Deal

According to details shared by both organizations, METR will get access to pre-deployment models, internal evaluation tools and infrastructure, and insight into Anthropic's Responsible Scaling Policy in action.

That includes the ability to probe models for dangerous capabilities, such as cyber-offense, deception, autonomous replication, and assistance with weapons of mass destruction. METR will also be able to review how Anthropic itself measures risk, and whether its internal safeguards are sufficient before launch.

Crucially, the access is ongoing rather than a one-off audit. METR is expected to evaluate successive generations of Claude models, allowing it to track safety trends over time as capabilities advance.

Why Waiving Veto Power Matters

The most striking element of the deal is what Anthropic gave up: control. Anthropic has agreed not to block or edit METR's findings. METR can publish its evaluations independently, even if the conclusions are critical of Anthropic.

That waiver of veto power addresses one of the longest-running criticisms of voluntary AI safety testing — that labs cherry-pick friendly auditors and quietly bury negative results. By guaranteeing METR's independence, Anthropic is betting that credible outside validation will build more public trust than tightly managed disclosures.

METR, formerly known as the Model Evaluation and Threat Research organization, has built a reputation as one of the most rigorous third-party evaluators in the field. Known for its work on autonomous capabilities and agentic tasks, the nonprofit has previously worked with OpenAI, Anthropic, and Google DeepMind, but never with this level of insider privilege.

What This Means for AI Accountability

For the broader AI industry, the Anthropic-METR pact could reshape how oversight works. Policymakers in the U.S., U.K., and EU have been pushing for mandatory third-party evaluations and stronger auditing regimes for frontier models. Until now, progress has been slow, hampered by concerns over trade secrets, safety risks from leaking model weights, and lack of auditor access.

This deal offers a working template: deep access paired with strict information security, plus guaranteed publication freedom.

Safety advocates are already calling on other labs to follow suit. If OpenAI, Google DeepMind, xAI, and Meta adopted similar employee-level auditing, it would mark a major shift from self-regulation toward true external accountability.

Challenges and Open Questions

Despite the optimism, major questions remain. How will METR handle highly sensitive information about vulnerabilities without inadvertently disclosing it? How will it be funded to sustain long-term deep evaluations? And will independent findings actually change Anthropic's deployment decisions if a model is found to be risky?

Both organizations say safeguards are in place to prevent leaks of proprietary model details while still allowing substantive public reporting. METR will disclose methodologies and high-level results, not weights or exploitable specifics.

Anthropic says it views the partnership as complementary to government safety institutes, not a replacement. The U.S. AI Safety Institute and U.K. AI Safety Institute will continue their own pre-deployment testing, but METR's insider position could provide faster, more continuous feedback.

A Blueprint for the Industry?

In an era defined by rapid capability jumps and growing public anxiety over AI risk, transparency has become a competitive differentiator. Anthropic, which was founded on a safety-first mission and has long published its scaling and alignment research, is now leveraging radical transparency as strategy.

Whether the gamble pays off will depend on what METR finds — and what Anthropic does next. If future Claude systems pass tough independent scrutiny, Anthropic will have set a new bar for trust. If they don't, the world will be watching how the company responds.

Either way, the era of black-box frontier AI development just got a little more open.

AndroGuider Team

Articles written by the AndroGuider team. We try to make them thorough and informational while being easy to read.